On October 4, 2026, read at 06:40 UTC, ten clouds listed the NVIDIA T4 and its median rate card stood at $0.56 per GPU-hour. The eight on-demand rates ran from $0.23 at immers.cloud to $0.81 at Replicate, 3.5x apart for the same 16 GB card, and the middle half of the ten rate cards sat between $0.43 and $0.62.
The eight on-demand rates
immers.cloud printed the lowest of them at $0.23, listed in roubles and converted at the Bank of Russia’s published rate of 0.011978 for October 3, 2026. Google Cloud followed at $0.35 for a T4 in Columbus, a GPU-only figure that does not carry the instance around it. Hugging Face listed $0.40 and $0.50. Amazon Web Services and Microsoft Azure printed the same number to the cent, $0.53, for g4dn.xlarge in us-east-1 and Standard_NC4as_T4_v3 in northcentralus. Baseten was at $0.63, Replicate at $0.81.
Three more clouds list the T4 as serverless capacity and sit in a separate table on the card page: Cerebrium and Modal at $0.59, Inferless at $0.66. They are the other three of the ten providers behind the $0.56 median.
One vendor, two of its own tables
Of the ten providers, Hugging Face is the only one we read at two different T4 prices. Its Spaces Hardware table prices “Nvidia T4 - small” at $0.40 an hour; further down the same page, under Inference Endpoints, “NVIDIA T4 x1 (aws)” is $0.50. The higher is 1.25x the lower ($0.50 / $0.40).
October 3 is the day we first read the Spaces Hardware table, not a day Hugging Face rewrote a price: the $0.40 and the $0.50 were both on that page when our parser was widened to read the second table, and the card page records the $0.40 as listed since October 3 for that reason. The same reading gave the A10G a third listing, at $1.00, against $1.01 at Amazon Web Services and $1.21 at Baseten, which put the A10G median rate card at $1.01 across three providers.
Spot and committed rates
One spot rate was live for the T4: $0.07 on Azure in australiacentral2, the cheapest of the 31 Azure regions we read. All four committed rates came from the same two hyperscalers. Google Cloud’s three-year commitment was $0.16 and Azure’s $0.20. Of the two one-year rates, Google Cloud’s was $0.22 and Azure’s $0.31, which the card page marks as $0.22 below that provider’s own on-demand $0.53 for the identical instance type. Azure’s one-year figure is the one committed rate of the four that stands above immers.cloud’s on-demand $0.23.
Nothing here is forecast, and none of it is advice on where to rent. These are the figures the ten pages carried at 06:40 UTC on October 4, 2026, and a rate is a listed price, not a promise of a free machine.
Checking these figures yourself
Every rate above links to the page it was read from, and /gpu/t4 carries all
sixteen readings with the capture time against each one: eight on-demand, one
spot, three serverless and four committed. The median rate card and the
$0.43-$0.62 middle half are the figures that page prints at the top, over one
rate card per provider. /prices.json holds the same rows with their
scraped_at stamps under CC BY, and the Bank of Russia rate behind the
immers.cloud conversion is printed beside that row with its date.